Skip to content

Dataset · mlfoundations

Trillion-token multimodal corpus with PDF and ArXiv

mlfoundations/MINT-1T-PDF-CC-2024-18

A massive open dataset of interleaved English text and images for pretraining multimodal models.

What's inside

The dataset contains interleaved English text and visual data: the full MINT-1T corpus includes 1 trillion text tokens and 3.4 billion images (this split's size ranges from 100B to 1T records). Beyond standard web sources, it includes previously unused materials from PDF documents and ArXiv research papers. Compiled by a team of researchers from the University of Washington.

What you can do with it

  • Pretraining large multimodal models from scratch
  • Training systems to understand complex documents, PDFs, and illustrated research papers
  • Training image-to-text generation models
Tasks
Image → text, Text generation
Languages
English
Size
100B–1T rows
Downloads this month
41.6K
Likes
32

How to load it

from datasets import load_dataset

ds = load_dataset("mlfoundations/MINT-1T-PDF-CC-2024-18")